Dynamic Prediction of Rectal Cancer Relapse and Mortality Using a Landmarking-Based Machine Learning Model: A Multicenter Retrospective Study from the Italian Society of Surgical Oncology—Colorectal Cancer Network Collaborative Group

Rossella Reddavid, Ugo Elmore, Jacopo Moro, Paola De Nardi, Alberto Biondi, Roberto Persiani, Leonardo Solaini, Donato P. Pafundi, Desiree Cianflocca, Diego Sasia, Marco Milone, Giulia Turri, Michela Mineccia, Francesca Pecchini, Gaetano Gallo, Daniela Rega, Simona Gili, Fabio Maiello, Andrea Barberis, Federico Costanzo, Monica Ortenzi, Andrea Divizia, Caterina Foppa, Gabriele Anania, Antonino Spinelli, Giuseppe S. Sica, Mario Guerrieri, Roberto Polastri, Francesco Bianco, Paolo Delrio, Giuseppe Sammarco, Micaela Piccoli, Alessandro Ferrero, Corrado Pedrazzani, Michele Manigrasso, Felice Borghi, Claudio Coco, Davide Cavaliere, Domenico D’Ugo, Riccardo Rosati, Danila Azzolina

Background: Almost 30% of patients with rectal cancer (RC) who submit to comprehensive treatment experience relapse. Surveillance plays a leading role in early detection. The landmark approach provides a more flexible and dynamic framework for survival prediction. Objective: This large retrospective study aims to develop a machine learning algorithm to profile the patient prognosis, especially the risk and the onset of RC relapse after curative resection. Methods: A cohort of 2450 RC patients were analyzed using landmark analysis. Model A applied a classical cause-specific Cox approach with a landmarking approach, while Model B implemented a landmarking-based RSF (random survival forest) competing risk algorithm. The two models were compared in terms of predictive and interpretative ability. A bootstrapped validation strategy was employed to validate the model’s performance and prevent overfitting. The best-performing hyperparameters were selected systematically, ensuring the model’s robustness within the landmark approach. The study assessed these factors’ importance and interactions using RSF and compared the predictive accuracy to that of the classical Cox model. Results: Model B outperformed Model A (mean C-index 0.95 vs. 0.78), capturing complex interactions and providing dynamic, individualized relapse predictions. Clinical factors influencing survival outcomes were identified across time with the landmark approach allowing for more accurate and timely predictions. Conclusions: The landmark approach offers an improvement over traditional methods in survival analysis. By accommodating time-dependent variables and the evolving nature of patient data, this approach provides a precise tool for profiling RC survival, thereby supporting more informed and dynamic clinical decision-making.